Building Your Own AI Platform and Tools Using ChatGPT

The talk opens with the speaker's experience pushing ChatGPT past its safety behavior, drawing a comparison to earlier elicitation work. It covers a range of prompt techniques for getting around refusals, including role-play framing, format and length constraints, regenerating responses, open-ended questions, and chaining follow-up questions so that each request looks harmless on its own.

The second half turns to building a small machine learning model for red team tasks. The workflow starts with defining an objective, then choosing a model type, gathering and cleaning data, generating synthetic examples, engineering features, and training with hyperparameter search. The demonstrations use scikit-learn on a laptop, including a question-answer classifier, a service version detector trained on packet capture features, and a rough attack severity predictor built on a very small dataset.

The speaker stresses that data preparation takes most of the effort and that the process is iterative, with results guiding the addition of data and features. The practical takeaway is how generative AI tools can speed up research and tooling work, along with the caveat that outputs need checking and that the demos were early experiments.